New Delhi: Amazon Web Services is advancing its AI hardware roadmap. At AWS re: Invent 2025, the company said its upcoming Trainium4 chips will integrate Nvidia’s NVLink Fusion interconnect, reflecting a strategic shift toward higher-bandwidth, tightly connected AI training systems. While AWS did not disclose a launch timeline, it positioned NVLink Fusion as a core element of its next-generation silicon, aimed at supporting larger and more complex AI models.
The move places AWS alongside Intel and Qualcomm in adopting NVLink, strengthening Nvidia’s influence across the AI infrastructure ecosystem. As part of the collaboration, AWS will also offer AI Factories—dedicated compute clusters deployed within customer data centers to support model training on sensitive or high-volume datasets.
AWS introduced new Trainium3-based servers as well, now available for customers. Each server contains 144 Trainium3 chips and delivers significantly higher compute performance while reducing power consumption compared to earlier AWS AI hardware. Executives said the company aims to compete on price efficiency and performance as enterprises evaluate alternatives beyond GPU-centric architectures.
AWS also updated its Nova foundation models, including the multimodal Nova 2 and the real-time speech model Sonic. A new service, Nova Forge, enables enterprises to train domain-specific models on their own data. The updates signal AWS’s broader effort to strengthen its position in the increasingly competitive AI cloud market.
